Lesson 18 · Chips & LLMs

EDA: The Duopoly That Taxes Every Chip

Cadence and Synopsys — the design-software toll booth that every chip ever made must pass through, including NVIDIA's, AMD's, and every custom ASIC.

Builds on: L1 (value chain), L10 (ASIC), L17 (ASML) Confirms: THESIS.md — Cadence/Synopsys (last name)

In your "one dollar of capex" trace, two names appeared invisibly upstream of everything: ASML, which prints every leading-edge chip, and the EDA duo, on whose software every chip is designed. You've now done ASML. This lesson finishes the pair — and with it, all seven names in your THESIS.md are grounded.

Core thesis: A modern chip has tens of billions of transistors — no human can lay that out by hand. Every chip is therefore compiled from human-readable design code into a fab-ready layout by EDA software, and that software is a two-player oligopoly (Cadence + Synopsys, with Siemens EDA third). The moat is switching costs so deep they're almost absolute: certified foundry flows, decades of accumulated IP, and engineers trained on one toolchain. EDA is the purest picks-and-shovels business in the stack — it gets paid on every design regardless of who wins any socket.

01 — What EDA Actually Does

Compiling a Chip: RTL → GDSII

Think of chip design as a compiler pipeline. An engineer writes behavior in a hardware description language; EDA tools transform it, step by step, into a physical layout the fab can print. Each stage is a distinct, expensive tool category — and Cadence and Synopsys sell across all of them.

StageWhat happensRepresentative tools
1. RTL design Engineer writes behavior in Verilog/VHDL (Register-Transfer Level) — "what the logic does," not where it sits. (human-authored HDL)
2. Logic synthesis RTL is compiled into a gate-level netlist — actual logic gates, optimized for timing/area/power. Synopsys Fusion Compiler · Cadence Genus
3. Verification The largest cost — simulation, formal proof, timing. Catching a bug here vs. after a $30M+ mask set is the whole game. Synopsys VCS · Cadence Xcelium / JasperGold
4. Place & route Physical design: place billions of gates and route the wiring between them, honoring timing and the foundry's rules. Cadence Innovus · Synopsys IC Compiler II
5. Signoff Final checks against the foundry rulebook: static timing (STA), power, DRC, LVS. Must pass exactly or the fab won't accept it. Synopsys PrimeTime · Cadence Tempus
6. Tape-out Output the final layout file — GDSII/OASIS — and send it to the fab. This is "tape-out." → to TSMC / Samsung / Intel

Every accelerator you've studied — NVIDIA's B200, AMD's MI350X, Google's TPU, Broadcom's custom ASICs — went through this exact pipeline. Broadcom's whole ASIC business (L10) is, in large part, EDA expertise applied for hire.

02 — The Duopoly and Its Switching Costs

Why It's a Two-Player Game

PlayerPositionStructure
SynopsysCo-leader; strongest in synthesis, signoff, and (uniquely) a huge IP businessDuopoly
CadenceCo-leader; strongest in place-and-route, custom/analog, and system designDuopoly
Siemens EDA (ex-Mentor)Clear #3; strong in verification & specific niches#3

The three control the overwhelming majority of EDA; Cadence and Synopsys are the two that span the entire flow at the leading edge. The structure is durable because the switching costs are extraordinary:

Connect to L17: EDA and ASML are the two structural monopolies/duopolies at the foot of the chain. Both get paid on every chip regardless of who wins downstream — but note the difference: ASML's moat is physics + supply-chain integration; EDA's is software switching costs + foundry co-certification. Different mechanism, similar durability.

03 — Two Growth Engines: AI Design and IP

Beyond the Core Toll Booth

AI-assisted design. Place-and-route is a brutal optimization problem; both leaders now use reinforcement learning to explore the design space automatically — Synopsys DSO.ai and Cadence Cerebrus. These deliver real productivity gains (better power/performance/area in less engineer-time), open a new monetization tier, and — crucially — deepen the moat by making the incumbent flow even more capable and harder to leave.

IP licensing. Synopsys in particular sells silicon-proven interface IP — the pre-built blocks (PCIe, USB, DDR/HBM PHY, etc.) that chip designers license rather than build. As chips integrate more standard interfaces (and HBM/PCIe generations churn fast), this is a structural growth vector. (Arm sells CPU-core IP; Synopsys sells the surrounding interface IP — complementary.) Synopsys's pending move into multiphysics simulation (the Ansys acquisition) extends the tool footprint further up the design process.

04 — Investment Implications

The Highest-Quality Business Model in the Stack?

Picks-and-shovels, socket-agnostic
EDA is paid on every design — NVIDIA, AMD, Google TPU, every Broadcom ASIC, every hyperscaler chip. The custom-silicon boom (L10) is a tailwind, not a threat: more design teams = more seats. You don't need to pick the accelerator winner to own the toll booth.
Recurring, high-margin, predictable
Subscription-heavy revenue + deep switching costs = low churn, high gross margin, and earnings visibility a hardware vendor can only envy. The least cyclical way to own semiconductor growth. Watch backlog and renewal rates.
Growth = complexity + design starts + AI/IP
Revenue scales with chip complexity (more transistors → more tool-intensity) and the number of design starts, now boosted by AI-design upsell and IP. Tie the growth read to overall R&D budgets and design-start activity, not unit shipments.
The bear case: maturity, China, open-source tail
Core market growth is tied to industry R&D (mature). Advanced-node EDA is export-restricted to China — a revenue cap and policy risk. Open-source flows (OpenROAD) are a long-tail threat but nowhere near production parity at advanced nodes today.
Thesis Update

Cadence/Synopsys — Now Grounded (★★★)

Key technical assumption: EDA switching costs and the two-player structure are durable because every advanced node ships with co-certified incumbent tool support — no third entrant or open-source flow reaches production parity at the leading edge. Falsifier — what would change your mind:

Primary Source

Go Deeper

Read first: Semiconductor Engineering — EDA / Design knowledge center — high-trust, engineer-written explainers on each stage of the flow and where the tools fit. Pair with the SemiWiki EDA coverage (in RESOURCES.md) for industry-structure context.

For the business model: read the latest Synopsys and Cadence investor decks side by side — note recurring-revenue %, backlog, and how each frames AI-design + IP as the growth story.

Comprehension Check

Quiz — 5 Questions

Select the best answer for each.

1. In the chip-design pipeline, what is "tape-out"?

Writing the initial RTL behavior in Verilog
Sending the final GDSII layout file to the fab
Running the first batch of wafers through lithography
Packaging the finished die with HBM

2. Which best describes why EDA is a "picks-and-shovels, socket-agnostic" business?

It only sells to NVIDIA, the market leader
Every chip design — GPU, ASIC, or hyperscaler silicon — must use it, so it wins regardless of which chip wins
It manufactures the chips itself, capturing fab margin
It licenses CPU cores like Arm does

3. What makes EDA switching costs so deep that the duopoly is stable?

Government licenses limit how many EDA firms can exist
Tools are co-certified to each foundry's PDK per node, plus years of accumulated flows, IP, and trained engineers
EDA software is given away free, locking in support contracts
Only one programming language can describe chips

4. The custom-ASIC boom (Broadcom, hyperscaler silicon) is, for EDA, best described as:

A threat that bypasses EDA tools entirely
A tailwind — more design teams means more tool seats and IP licenses
Irrelevant, since ASICs don't need place-and-route
A reason margins will collapse

5. Which event would most directly falsify the Cadence/Synopsys durability thesis?

A new HBM generation shipping from SK Hynix
An open-source or third-party flow winning production tape-outs at an advanced node
NVIDIA releasing a faster GPU than AMD
TSMC raising wafer prices
From your instructor: The one-liner — every chip is compiled from code to layout by software only two companies sell at the leading edge, and the moat is foundry-certified switching costs. With this, all seven THESIS names are grounded. Ask me anything: the synthesis-vs-P&R distinction, how DSO.ai actually optimizes, or how to read EDA recurring-revenue mix against the chip cycle.